A spectral based visual matching method for image classification

Visual matching algorithms can be described in terms of visual content representation and similarity measure. With local feature based representations, visual matching can be restated as: (1) how to obtain visual similarity from the local kernel matrix, and (2) how to calculate the local kernel matrix effectively and efficiently. Existing methods mostly focus on the former, and use Euclidean distance to calculate the local kernel under Gaussian noise assumption. However, this assumption may not be optimal for gradient based local features. In this paper, we propose a Local Coding based Spectral Analysis (LCSA) method to exploit the low dimensional manifold structure in feature space. Specifically, we select a set of anchor points, and represent each feature as a linear combination of anchor points with locality constraint. The spectral analysis can then be efficiently processed according to this representation. Following the derivation of Efficient Match Kernel (EMK) [6], a compact lower-dimensional set-level image representation is obtained for visual similarity measure. Experimental results on several benchmark image classification datasets, i.e. 15-scenes and Caltech101/256, show superior performance compared with the existing state-of-the-art techniques with SIFT feature.